Skip to content

Source paper

LoGeR: Long-Context Geometric Reconstruction with Hybrid Memory

Junyi Zhang, Charles Herrmann, Junhwa Hur and 5 more

Junyi Zhang, Charles Herrmann, Junhwa Hur, Chen Sun, Ming-Hsuan Yang, Forrester Cole, Trevor Darrell, Deqing Sun

arXiv 2603.03269 · 202643 upvotes on Hugging Face

Research ideas from this paper (3)

Generated by AI. The paper’s authors did not write these.

AI summary

This paper introduces LoGeR, a novel feedforward geometric foundation model designed for long-context dense 3D reconstruction. It overcomes the quadratic complexity of global attention and sequence length limitations by utilizing a chunk-wise processing architecture with a hybrid memory system, achieving state-of-the-art results on extremely long video sequences.

How it works (AI summary)

LoGeR processes video streams in chunks using a dual-component hybrid memory module. It combines non-parametric Sliding Window Attention (SWA) to preserve high-fidelity local geometry between adjacent chunks, and a parametric Test-Time Training (TTT) layer to compress and propagate long-range global context, which anchors the coordinate frame and prevents trajectory drift.